Kevin Niparko's open secret: most PMs do not actually manage anyone. They manage a product. Nobody reports to the mini-CEO. At SpaceXAI the product team ships code and prototypes in small teams, and for the first time he has a team that reports to him — of agents. Grok Bot is a double-digit percent of internal merged PRs. He used it for a few months ahead of launch to build the product.

The primitive that replaced the stale weekly priority list is the attention list. Chief of Staff watches Slack, email, Granola, and calendar and writes down what he is actually focused on: incident channels, a candidate close, a launch blog sitting in review. Agents use that list as a filter so the thousand Slack messages that are not the work do not become the work. He can also compare stated goals against where attention actually went.

Roster: Chief of Staff is the only generalist — calendar, Slack, inbox, the attention list. Stays quiet if nothing changed. Hourly it archives Gmail unless the thread names him and maps to the list. Last week it sent holds and asked him to pick a start time; he said 10am; it moved the hold. Eng mgr (Emily) does not code. She breaks work down, delegates to IC eng agents, and checks output against the goal. He onboarded her by having her read all of @danielstjules's Slacks so she would know what great looks like. Five eng agents underneath spin Cloud Agents, test, and coordinate. Data analyst (Ashley) is the one he uses most — warehouse, charts, tracking goals, dash every morning. Tactical questions that used to go unanswered because the friction was too high now come back as synthesized charts; the hard, ambiguous work still goes to the human DS team. PM Pete sits in the work: RFCs, research, feedback, docs, hardware. Example: turn a channel argument about a feature split into an RFC (P0–P2) with quotes. They hacked on a Raspberry Pi together; Pete researched a missing part, ordered it off Amazon, and taught him how to wire it. Recruiter finds undiscovered talent and runs loops.

Why not one omniscient agent: referenceability, parallelism, scoped memory. CoS should not be debugging computer-use evals. The eval agent is not archiving mail. Named agents, separate memory. Highest-leverage instruction he has found: read my past Slack messages and develop a Slack style guide. Agents bias to action and can over-communicate — tell them to stay quiet unless something needs him.

Approvals: he still sends most messages and edits most docs. External email, purchasing, deletes — human. Pete ordering a Pi part was a live exception he is explicit about; the standing rule is still his.

Official guide on x.ai, Aug 15, 2026.